EShuffleNetV2-IUNet: ensemble deep learning-based model for brain stroke classification and segmentation on computed tomography images

Background Considering the many features of tumors, brain stroke classification is a crucial topic for research in the field of medical imaging. Categorizing brain strokes from computed tomography (CT) images is one of these complex problems. Since classification accuracy is crucial, computer vision researchers are developing several methods, yet identifying brain strokes with high accuracy continues to take time. The patient’s chances of survival are increased when brain stroke types are diagnosed early, and treatment is initiated promptly. Methods Deep learning models have shown promise in recent years, particularly in helping neurologists classify brain strokes. The primary goal of the research is to use ensemble deep learning models to categorize brain CT scan images into hemorrhagic stroke, ischemic stroke, and normal. Initially, pre-processing involves applying three-stage image processing techniques, including noise reduction, image enhancement, and data augmentation. Further, pre-processed images are loaded into the newly created Enhanced ShuffleNet V2 (EShuffleNetV2) model for effective stroke classification. Then, the region of the stroke lesions is segmented using the Improved U-Net (IUNet) model. Brain stroke classification and segmentation are performed using publicly accessible brain CT imaging datasets. Results The suggested method correctly separates the lesion region with a 0.9668 Dice score and a 0.9792 Intersection over Union (IoU), and it successfully classifies the brain stroke images with the most excellent accuracy of 99.35%. According to our results analysis, our proposed ensemble deep learning system outperformed state-of-the-art brain stroke classification and segmentation methods. Unlike existing methods that only adapt conventional Convolutional Neural Networks (CNNs) or U-Net architectures with minor modifications, our approach introduces an Enhanced ShuffleNetV2 with a novel Parallel Mixed Attention Mechanism (PMAM) for robust feature refinement and an IUNet with Residual Extended Skip (RES) and Wide Context (WC) blocks for precise lesion segmentation. This unique ensemble framework ensures both computational efficiency and state-of-the-art accuracy, providing a new direction for lightweight yet powerful stroke classification and segmentation.